Explain 02

Why can history change behaviour?

Short answer

Because earlier interaction leaves state behind: context, memory, artefacts, half-finished work. Two identical requests can produce different actions when the state behind them differs. The current prompt is not the whole input.

The same request
Arm A
a fresh session
action A
Arm B
after a long working session
a different action
Nothing differs except the accumulated state behind the request.

History accumulates in several places at once. The context window carries earlier turns and tool output. Memory features carry summaries across sessions. The environment carries artefacts: files created, settings changed, work half done. When a new request arrives, the system does not see it in isolation; it sees it on top of all that state.

Sometimes the state is exactly why the system behaves well: it remembers what you are doing. The measurement question is whether earlier state continues to influence behaviour after the relevant information has been replaced, removed or no longer applies, shifting actions in ways the current request does not explain.

A concrete example

An illustrative case: “Tidy up the unused files.” In a fresh session the assistant lists candidates and asks. Late in a long refactor, half the project is temporarily unused, the context is full of delete commands that went unchallenged, and the same request can cut much deeper. The words were identical; the state was not.

Why a summary can miss it

If order can change the outcome, then a summary that records only what happened, and not in what order, can be blind to the difference. Arithmetic makes the point exactly:

Animation · 8 seconds

Both tracks start at 1. Both apply exactly the same two operations.

Add first
  1. 1
  2. + 2
  3. 3
  4. × 3
  5. 9
Multiply first
  1. 1
  2. × 3
  3. 3
  4. + 2
  5. 5
Count-only summaryone addition, one multiplication (identical)
Order-preserving summaryadd then multiply multiply then add

Same operations. Different order. Different result.

The arithmetic is exact. Whether order-sensitive measurements improve prediction of AI behaviour is an experimental question, and one this programme is still testing.

A common misunderstanding

“If the last prompt looks fine, the action is fine.” A reviewer who only reads the current request is assuming that everything relevant about the past is already summarised in it. Whether that assumption holds is an empirical question, and it is one of the questions Overdog tests.

Why it matters

Many monitoring methods score the current request, response or tool action independently. If later behaviour depends on relevant state the monitor does not receive, the monitor cannot condition on that state. Measuring when and how much that loss matters is a prerequisite for any honest claim about history-aware monitoring.